Add model_index.json for compatibility with Diffusion Pipeline
I think it still needs the diffusers based model in order to work with your app.
I'll look into uploading them.
Yes, I think you're right. Great, thanks. π
Hi
@nitrosocke
, awesome model, excited to try it out with diffusers
pipelines! This conversion script should do the trick, if your original checkpoint is compatible with CompVis/stable-diffusion
: https://github.com/huggingface/diffusers/blob/main/scripts/convert_original_stable_diffusion_to_diffusers.py
Thank you! I actually train on diffusers, so the ckpt file is the converted model.
I just need to get my git working with the huggingface hub and access token in order to upload the folder structure.
Hey @nitrosocke ,
π a maintainer of diffusers
here.
Great work on arcane stable diffusion! The model looks great, congrats!
I'm curious how we could improve our user experience to upload the folder structure & how to improve the user experience of huggingface_hub
and/or git, any feedback of what is not easy to understand is more then welcome! E.g. it seems like you managed well to upload the .ckpt
to the Hub, but no the folder structure for the diffusers
checkpoint - any reason why?
Hi
@patrickvonplaten
Thank you for your kind words and for looking into this.
I am not an experienced GitHub or HuggingFace user, so this might be not that difficult after all, but I wasn't able to upload my local folder structure to the existing model repo here.
I use Miniconda with python 3.9 on windows 10 with git and the huggingface hub installed in a condo environment.
When using the hf cli Login the access token gets saved successfully. When using the git lfs pull command after that it still pops up the old login fields with username and password and failes to validate those.
When using a python script with the method from the huggingface documentation it works but failed to upload to the repo. I will try again and can provide the errors it gave when it failed.
Uploading the ckpt file was very easy as HF has the option to select files from the Ui. There is no option for uploading folder structures or creating branches as far as I can tell.
Hope this gives you an idea of my workflow and thought process.
This makes a lot of sense - thanks for sharing! BTW, I converted the weights and opened a PR to add weights to diffusers
you can see here: https://huggingface.co/nitrosocke/Arcane-Diffusion/discussions/11
Given the interest in this model, maybe it could also make sense to open a GPU powered space for this
@akhaliq
?
On the Hub side we'll improve support to be able to drag'n'drop any folder (with nested structure) to upload it, directly from the website
cc @victor @coyotte508 πͺ
Thank you @julien-c for the update on that feature! Really looking forward to it as that would speed up the workflow immensely for a hub beginner like me.